Stop Asking ChatGPT If Your Idea Is Good: Run a Board Meeting

WatchInstagram

Ask ChatGPT whether your idea is good and it'll probably say yes. A Stanford study published in Science tested 11 leading AI models, including ChatGPT, Claude and Gemini, and found they backed the user's actions 49% more often than other people did.

So stop asking one friendly assistant. Make it hold a board meeting: five advisors whose only job is to find the holes in your idea, a blind review of their answers, then one verdict and one next move.

This page has the board prompt to paste into ChatGPT, Claude or Gemini, free, with nothing to install. After it come two council tools: one for Claude Code, and one that makes several AI models argue it out.

How do you run an AI board of advisors in ChatGPT?

Paste the prompt below into ChatGPT, Claude or Gemini and fill in the three lines at the top: your decision, the context, and what you're leaning towards. Then read the Contrarian's answer before you commit to anything.

It's built on Ole Lehmann's LLM Council roles. Each of the five advisors has to push back on at least one part of your plan, a blind review compares their answers, and the chair ends with one thing to do in the next 48 hours.

The board prompt

You're chairing a board meeting about a decision I'm about to make. Don't give me your own opinion and don't reassure me. Run the meeting exactly like this.

MY DECISION: [what you're deciding, e.g. raise my prices 20%, hire my first employee, take the job offer]
CONTEXT: [the numbers, the deadline, what's at stake, what you've already tried]
WHAT I'M LEANING TOWARDS: [your current answer, and why]

STEP 1: FIVE ADVISORS ANSWER
Each advisor answers on their own, as if they can't see the others. Each one has to push back on at least one part of my plan.
1. The Contrarian: assumes this has a fatal flaw and hunts for it. What kills this?
2. The First-Principles Thinker: ignores how I framed the question, finds the real problem underneath, and asks whether I'm solving the right one.
3. The Upside Hunter: looks for the bigger opportunity, or the option I haven't even listed.
4. The Outsider: knows nothing about me or my field and reacts only to what's in front of them. What's confusing? What sounds off?
5. The Executor: asks what I'd actually do on Monday morning, and flags anything with no clear first step.

STEP 2: BLIND REVIEW
Shuffle the five answers and relabel them A to E with no names. Then, as an independent reviewer, say which answer is strongest and why, the biggest blind spot in each, and what all five missed.

STEP 3: THE CHAIR'S VERDICT
As the chair, give me: where the board agrees, where it splits (keep the strongest dissent, don't smooth it over), a clear recommendation (go, don't go, or go with changes), and ONE first action I can take in the next 48 hours.

Rules: keep each advisor to about 120 words. Be blunt, no flattery, no "great question". If something you'd need to decide is missing, say what it is and what you're assuming, then run the meeting anyway.

If you use Claude Code or Cowork, Ole Lehmann's own LLM Council skill runs the same five advisors as sub-agents and writes up a report. Add it to your skills, then type "council this" and your question.

What are the other two ways to run an AI council?

claude-council: a five-advisor debate inside Claude Code

claude-council is a free Claude Code skill built on the same idea: five advisors, anonymised peer review, and a journal that tracks how your decisions turned out. When the reviewers agree too easily, it forces a debate before you get your answer. Quick mode runs three advisors in about 30 seconds, and Deep mode always debates.

It's brand new (version 0.1.0) and needs Claude Code on a Pro plan or higher, plus jq for the journal.

npx skills add TorpedoD/claude-council

Then type /claude-council and your decision. Add quick or deep to change the mode.

LLM Council: several AI models grade each other blind

Andrej Karpathy, an OpenAI co-founder, built LLM Council as a Saturday hack, and it has over 25,000 stars on GitHub. Instead of trusting one model, it sends your question to several: GPT, Gemini, Claude and Grok by default. Each one ranks the others' answers without knowing whose is whose, then a chairman model writes the final reply.

It runs on your own computer through OpenRouter, so you need an OpenRouter API key with paid credits. Karpathy says he won't support or improve it, so treat it as a starting point rather than a product.

A board meeting won't make the decision for you, but it'll find the hole before your customers do. To see where AI could save your team the most time, take the free AI bottleneck test, or read our insights on AI implementation.